AYOLOv2

The main goal of this repository is to rewrite the object detection pipeline with a better code structure for better portability and adaptability to apply new experimental methods. The object detection pipeline is based on Ultralytics YOLOv5.

What's inside of this repository

YOLOv5 based portable model (model built with kindle)
Model conversion (TorchScript, ONNX, TensorRT) support
Tensor decomposition model with pruning optimization
Stochastic Weight Averaging(SWA) support
Auto search for NMS parameter optimization
W&B support with model save and load functionality
Representative Learning (Experimental)
Distillation via soft teacher method (Experimental)
C++ inference (WIP)
AutoML - searching efficient architecture for the given dataset(incoming!)

How to start
Pretrained models
Advanced usages
References
Contributors ✨

How to start

Install

Conda virtual environment or docker is required to setup the environment

Using conda environment

git clone https://github.com/j-marple-dev/AYolov2.git
cd AYolov2
./run_check.sh init
# Equivalent to
# conda env create -f environment.yml
# pre-commit install --hook-type pre-commit --hook-type pre-push

Using docker

Building a docker image

./run_docker.sh build
# You can add build options
# ./run_docker.sh build --no-cache

Running the container

This will mount current repository directory from local disk to docker image

./run_docker.sh run
# You can add running options
# ./run_docker.sh run -v $DATASET_PATH:/home/user/dataset

Executing the last running container

./run_docker.sh exec

Train a model

Example

python3 train.py --model $MODEL_CONFIG_PATH --data $DATA_CONFIG_PATH --cfg $TRAIN_CONFIG_PATH
# i.e.
# python3 train.py --model res/configs/model/yolov5s.yaml --data res/configs/data/coco.yaml --cfg res/configs/cfg/train_config.yaml
# Logging and upload trained weights to W&B
# python3 train.py --model res/configs/model/yolov5s.yaml --wlog --wlog_name yolov5s

Prepare dataset

Dataset config file

train_path: "DATASET_ROOT/images/train"
val_path: "DATASET_ROOT/images/val"

# Classes
nc: 10  # number of classes
dataset: "DATASET_NAME"
names: ['person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus', 'train', 'truck', 'boat', 'traffic light']  # class names

Dataset directory structure
- One of labels or segments directory must exist.
- Training label type(labels or segments) will be specified in the training config.
- images and labels or segments must have a matching filename with .txt extension.

DATASET_ROOT
│
├── images
│   ├── train
│   └── val
├── labels
│   ├── train
│   └── val
├── segments
│   ├── train
│   └── val

Training config

Default training configurations are defined in train_config.yaml.
You may want to change batch_size, epochs, device, workers, label_type along with your model, dataset, and training hardware.
Be cautious to change other parameters. It may affect training results.

Model config

Model is defined by yaml file with kindle
Please refer to https://github.com/JeiKeiLim/kindle

Multi-GPU training

Please use torch.distributed.run module for multi-GPU Training

python3 -m torch.distributed.run --nproc_per_node $N_GPU train.py --model $MODEL_CONFIG_PATH --data $DATA_CONFIG_PATH --cfg $TRAIN_CONFIG_PATH

- N_GPU: Number of GPU to use

Run a model validation

Validate from local weights

python3 val.py --weights $WEIGHT_PATH --data-cfg $DATA_CONFIG_PATH

You can pass W&B path to the weights argument.

python3 val.py --weights j-marple/AYolov2/179awdd1 --data-cfg $DATA_CONFIG_PATH

TTA (Test Time Augmentation)
- Simply pass --tta argument with --tta-cfg path
- Default TTA configs are located in res/configs/cfg/tta.yaml

python3 val.py --weights $WEIGHT_PATH --data-cfg $DATA_CONFIG_PATH --tta --tta-cfg $TTA_CFG_PATH

Validate with pycocotools (Only for COCO val2017 images) Future work: The val.py and val2.py should be merged together.

python3 val2.py --weights $WEIGHT_PATH --data $VAL_IMAGE_PATH --json-path $JSON_FILE_PATH

Pretrained models

Name	W&B URL	img_size	mAP^val 0.5:0.95	mAP^val 0.5	params
YOLOv5s	_{j-marple/AYolov2/33cxs5tn}	640	38.2	57.5	7,235,389
YOLOv5m	_{j-marple/AYolov2/2ktlek75}	640	45.0	63.9	21,190,557
YOLOv5l decomposed	_{j-marple/AYolov2/30t7wh1x}	640	46.9	65.6	26,855,105
YOLOv5l	_{j-marple/AYolov2/1beuv3fd}	640	48.0	66.6	46,563,709
YOLOv5x decomposed	_{j-marple/AYolov2/1gxaqgk4}	640	49.2	67.6	51,512,570
YOLOv5x	_{j-marple/AYolov2/1gxaqgk4}	640	49.6	68.1	86,749,405

Advanced usages

Export model to TorchScript, ONNX, TensorRT

You can export a trained model to TorchScript, ONNX, or TensorRT
INT8 quantization is currently not supported (coming soon).
Usage

python3 export.py --weights $WEIGHT_PATH --type [torchscript, ts, onnx, tensorrt, trt] --dtype [fp32, fp16, int8]

Above command will generate both model and model config file.

Example) FP16, Batch size 8, Image size 640x640, TensorRT

model_fp16_8_640_640.trt
model_fp16_8_640_640_trt.yaml

batch_size: 8
conf_t: 0.001  # NMS confidence threshold
dst: exp/  # Model location
dtype: fp16  # Data type
gpu_mem: 6  # GPU memory restriction
img_height: 640
img_width: 640
iou_t: 0.65  # NMS IoU threshold
keep_top_k: 100  # NMS top k parameter
model_cfg: res/configs/model/yolov5x.yaml  # Base model config location
opset: 11  # ONNX opset version
rect: false  # Rectangular inference mode
stride_size: 32  # Model stride size
top_k: 512  # Pre-NMS top k parameter
type: trt  # Model type
verbose: 1  # Verbosity level
weights: ./exp/yolov5x.pt  # Base model weight file location

Once, model has been exported, you can run val.py with the exported model.
- ONNX inference is currently not supported.
```
python3 val.py --weights model_fp16_8_640_640.trt --data-cfg $DATA_CONFIG_PATH
```

Applying tensor decomposition

A trained model can be compressed via tensor decomposition.
Decomposed conv is composed of 3 convolutions from 1 large convolution.
- Example)
  - Original conv: 64x128x3x3
  - Decomposed conv: 64x32x1x1 -> 32x16x3x3 -> 16x128x1x1

Usage

python3 decompose_model.py --weights $WEIGHT_PATH --loss-thr $DECOMPOSE_LOSS_THRESHOLD --prune-step $PRUNING_STEP --data-cfg $DATA_CONFIG_PATH

...
[  Original] # param: 86,749,405, mAP0.5: 0.678784398716757, Speed(pre-process, inference, NMS): 0.030, 21.180, 4.223
[Decomposed] # param: 49,508,630, mAP0.5: 0.6707606125947304, Speed(pre-process, inference, NMS): 0.030, 20.274, 4.345
Decomposition config saved to exp/decompose/val/2021_0000_runs/args.yaml
Decomposed model saved to exp/decompose/val/2021_0000_runs/yolov5x_decomposed.pt

Passing prune-step to 0 will skip pruning optimization.

Summary of tensor decomposition process

Pass random tensor x to original conv (ŷ) and decomposed conv (ỹ)
Compute E = Error(ŷ, ỹ)
If E < loss-thr, use decomposed conv
Apply pruning ratio with binary search
Jump to 1 until differential of pruning ratio is less than prune-step

:: Note :: Decomposition process uses CPU only.

Knowledge distillation

An ad-hoc implementation of the knowledge distillation motivated from the method in "End-to-end semi-supervised object dection with soft teacher".
Create pseudo-labels for "unlabeled dataset" using the inference of the "teacher" model.
:: Note ::
- Implemented to use the same dataset for the "training dataset" and the "unlabeled dataset". To use different datasets, the creation of the dataloader instance unlabeled_loader in distillation.py should be modified.
- Teacher model weights are fixed during training student model. (In the original paper, teacher model is updated using "exponential moving averaging" the student model.)

Usage

python distillation.py --model res/configs/model/yolov5s.yaml \
                       --cfg res/configs/cfg/distillation.yam \
                       --data res/configs/data/coco.yaml \
                       --teacher {wandb_runpath_of_pretrained_model}

Representation learning

Representations of a model can be automatically discovered from raw data by representation learning.
You can apply SimpleRL or SimCLR to find better representations of the model with --rl-type option.
- SimpleRL is a method to minimize a difference between last two representations of a model with L1 loss.
- SimCLR is a simple framework for contrastive self-supervised learning of visual representations without requiring specialized architectures.

Usage (default: base)

SimpleRL

python train_repr.py --model res/configs/model/yolov5s_repr.yaml \
                     --data res/configs/data/coco_repr.yaml \
                     --cfg res/configs/cfg/train_config_repr.yaml \
                     --rl-type base

SimCLR

python train_repr.py --model res/configs/model/simclr.yaml \
                     --data res/configs/data/coco_repr.yaml \
                     --cfg res/configs/cfg/train_config_simclr.yaml \
                     --rl-type simclr

Auto search for NMS parameters

If want to optimize NMS parameters(IoU threshold and confidence threshold), there are two ways to optimize.

Why there are two ways?

There is an issue with YOLOv5 validation.
It's ok with training or validating but the validation results are little different.

Optimize parameters with YOLOv5 validation.
Optimize parameters with COCO validation (pycocotools).

1. Optimize parameters with YOLOv5 validation.

python3 val_optimizer.py --weights ${WEIGHT_PATH | WANDB_PATH} --data-cfg $DATA_CONFIG_PATH

2. Optimize parameters with COCO validation.

python3 val_optimizer.py --weights ${WEIGHT_PATH | WANDB_PATH} --data-cfg $DATA_CONFIG_PATH --run-json --json-path $JSON_FILE_PATH

The --json-path is optional.

Advanced usage

If you have baseline network, give --base-map50 and --base-time arguments which are used for objective function.
To avoid the optimized parameters overfits, use --n-skip option to skip some images.

Applying SWA(Stochastic Weight Averaging)

There are three steps to apply SWA (Stochastic Weight Averaging):

Fine-tune pre-trained model
Create SWA model
Test SWA model

1. Fine-tune pre-trained model

Example

$ python train.py --model yolov5l_kindle.pt \
                  --data res/configs/data/coco.yaml \
                  --cfg res/configs/cfg/finetune.yaml \
                  --wlog --wlog_name yolov5l_swa \
                  --use_swa

2. Create SWA model

Example

$ python create_swa_model.py --model_dir exp/train/2021_1104_runs/weights \
                             --swa_model_name swa_best5.pt \
                             --best_num 5

Usage

$ python create_swa_model.py --help
usage: create_swa_model.py [-h] --model_dir MODEL_DIR
                           [--swa_model_name SWA_MODEL_NAME]
                           [--best_num BEST_NUM]

optional arguments:
  -h, --help            show this help message and exit
  --model_dir MODEL_DIR
                        directory of trained models to apply SWA (default: )
  --swa_model_name SWA_MODEL_NAME
                        file name of SWA model (default: swa.pt)
  --best_num BEST_NUM   the number of trained models to apply SWA (default: 5)

3. Test SWA model

Example

$ python val.py --weights exp/train/2021_1104_runs/weights/swa_best5.pt \
                --model-cfg '' \
                --data-cfg res/configs/data/coco.yaml \
                --conf-t 0.1 --iou-t 0.2

References

Object Detection

[1] Ultralytics YOLOv5 - https://github.com/ultralytics/yolov5

[2] YOLOR implementation - https://github.com/WongKinYiu/yolor.git

[3] MobileViT implementation - https://github.com/chinhsuanwu/mobilevit-pytorch

[4] Kindle - Making a PyTorch model easier than ever! - https://github.com/JeiKeiLim/kindle

[5] Wang, Chien-Yao, I-Hau Yeh, and Hong-Yuan Mark Liao. "You Only Learn One Representation: Unified Network for Multiple Tasks." arXiv preprint arXiv:2105.04206 (2021).

[6] Mehta, Sachin, and Mohammad Rastegari. "MobileViT: Light-weight, General-purpose, and Mobile-friendly Vision Transformer." arXiv preprint arXiv:2110.02178 (2021).

[7] Ghiasi, Golnaz, et al. "Simple copy-paste is a strong data augmentation method for instance segmentation." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2021.

Stochastic Weight Averaging

[8] SWA Object Detection implementation - https://github.com/hyz-xmaster/swa_object_detection

[9] Izmailov, Pavel, et al. "Averaging weights leads to wider optima and better generalization." arXiv preprint arXiv:1803.05407 (2018).

[10] Zhang, Haoyang, et al. "Swa object detection." arXiv preprint arXiv:2012.12645 (2020).

Knowledge Distillation

[11] Xu, Mengde, et al. "End-to-End Semi-Supervised Object Detection with Soft Teacher." arXiv preprint arXiv:2106.09018 (2021).

[12] He, Kaiming, et al. "Momentum contrast for unsupervised visual representation learning." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2020.

[13] Chen, Ting, et al. "A simple framework for contrastive learning of visual representations." International conference on machine learning. PMLR, 2020.

[14] Grill, Jean-Bastien, et al. "Bootstrap your own latent: A new approach to self-supervised learning." arXiv preprint arXiv:2006.07733 (2020).

[15] Roh, Byungseok, et al. "Spatially consistent representation learning." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2021.

Tensor Decomposition and Pruning

[16] PyTorch tensor decompositions - https://github.com/jacobgil/pytorch-tensor-decompositions

[17] PyTorch pruning tutorial - https://pytorch.org/tutorials/intermediate/pruning_tutorial.html

Representation Learning

[18] Bengio, Yoshua et al. "Representation Learning: A Review and New Perspectives." IEEE Transactions on Pattern Analysis and Machine Intelligence. 2013.

[19] Chen, Ting et al. "A Simple Framework for Contrastive Learning of Visual Representations." Proceedings of the 37th International Conference on Machine Learning. 2020